Machine Learning, Clustering, and Polymorphy

I I

Machine Learning, Clustering and Polymorphy

I

Stephen Jose Hanson

I

Bell Communications Research

I Abstract

I I

(WITT)

This paper reports a machine induction program prototypical

which attempts to model human

Properties of categories that human subjects are sensitive to include, best or

categorization.

members,

relative

contrasts

between

putative

categories,

and

polymorphy

(niether ne�ary or sufficient features). This approach represents an alternative to traditional

I

Artificial Intelligence (AI) approaches to generalization and conceptual clustering which tend

to focus on nec�ry and sufficient feature rules, equivalence classes, and search and match algorithms. The present approach is shown to be more consistent with human categorization

I

while

potentially

including

results

produced

by

more

traditional

clustering

schemes.

Applications of this categorization approach are also discussed in the domains of Expert systems and Information retrieval.

I I

Introduction

Most

current

Intelligence

I I

assumptions are consistent with human categorization data.

on

work

done

machine

in

Artificial

learning

and

conceptual clustering--and for that matter

most generalization schemes that have been

proposed

in Al-typically rest on five false

premises:

I I I I I

contrast,

psychological literature

results

are

in

the

inconsistent

with each of the five premises above. People

do not seem to try to form categories by determining the necessary and sufficient set

(1) that necessary and sufficient

I

In

categorization

feature lists must be central to the categorization engine;

(2)

that categories equivalence classes;

are

of "defining features" (Michalski,

(5)

the

and

that

to

minimize while

"variance"

maximizing

The distinction drawn here is somewhat subtle and

sufficient features, that would imply that people could not use common features w hich is contrary to intuition. However, there are many Jn'SSible mechanisms for achieving necessar y & sufficient categories, as exemplified in the "contrast approach" advocated below. Nonetheless, such categories for the prt:5ent a pproac h are a special case rather then a central purpose of the categor ization engi ne.

symbolic

top

tend

clusters

does not imply that people and animals do not have or know about categories that possess necessary and

(4) that probability measures are to

people

within

polynwrphy (neither necessary or sufficient features) rules are either uninteresting or noise;

antagonistic manipulation;

for

relative "contrasts" between categories; that

is,

1.

(3) that

1980)

a set of objects.1 Rather people seem to form

four 117

I (2)

"variance" between clusters (Rosch & Lloyd. 1978; Smith & Medin. 1981 ). People also tend to have best or prototypical members of a category as oppa;ed to equivalence cl�s CHoma. 1978; Posner & Keele,1968). Many categories that people use (perhaps all natural categories) have all or at least some members that possess neither necessary nor sufficient features and can best described by a polymorphy rule ("m features out of n", mV 1979 Mar Vol 39(9-B) 4632�633

'I># 21800: GROU P PROBLEM SOLVING 'It# 23510: HUMAN SEX DIFFERENCES 'I># 55520: VERBAL CXJMMUNJCATION 'I># 57230: WRIJTEN LANGUAGE 'I># 26250: rNTERPERSONAL rNTERACTION 'I># I0970: CXJMPU TERS 'I,# 29350: MAN MACHINE SYSTEMS abstract 7

'1.. PA VJ7 N 5 (1982)- No. 52137

I

'I>A Vi# 55515: VERBAL STIMUU

'I># 55990: V1SUAL STIMULATION 'I># 23480: HUMAN INFORMATION STORAGE '1.. PA V17 N 5 (1982)- No. 52142

abotract 6

'I>A Yio, Jun ll "'>T Visual recognition of words versus nonwords.. Cf.>J Disstrlltion Abstracts lnttnational

I I

'l>V 1979 Msr Vol 39(9-B) 4630

'I># 55981: VISUAL SEARCH

'I># 57020: WORDS (PHONETIC UNITS) 'I># 34340: NONSENSE SYLLABLE LEARNING 'I># 24420: ILLUMrNATION 'I># Jl.S60: CXJNnXTUAL ASSOClA TIONS 'I># 49220: SPELlrNG '1.. PA VJ7 N 5 (1982)- No. 52335

I

abstract 4

'I>A Richt er, Grecory 'l>T The relatior15hip between individual and developmental di6erenc.. in ocanning behsvior and clV 1981 May Vol 4JCJJ-B) 4287 'I># 01360: AGE DJFFI:C :REN ES

I I

'I># 45540: SCHOOL AGE CHJLDREN 'I># 00950: AOOLESCENTS

'I># 24 700: rNODENTAL LEARNING 'I># 55981: VISUAL SEARCH '1.. PA VJ7 N 5 ( 1982)- No. 52499 11\.A Korant. Leslie L 'I>T Elfects of t... o visual

trainin&

abstract 3

procrams upon automaticity of letter and "''Ot'd recoenition in urban Black kindercartners.

'I>J Dwertation Abstracts International

I

'IV 1981 Jun Voi4HI2-A. Pt 1) 4959

'I># 27 370: KINDERGARTEN STUDENTS 'I># 43350: RECXJGNITION (LEARNrNG) 'I># 51020: WORDS (PHONETIC UNITS) 'I># 282JO: LETTERS (ALPHABET)

I I

'I># 54940: URBAN ENVIRONMENTS �# 06150: B L ACKS 'I># 16190: EDUCATIONAL PROGRAMS 'l># 55980: VISUAL PERCEPTION

'1.. PA VJ7 N 5 (1982)- No. 52613

abstract 5

'I>A Pushka.sh. Mark 'l>T Elfect of the content of visually presented subliminal stimulation on oemantic and igural learning taU performance. �J Oi:\Sertation Abstracts lnttrnational "V 1981 Jun Vo14H 12 A. P1 JJ 5036

I

'l-# 55550: VERBAL LEARNING 'I# 34370: NONVERBAL LEARNING 'l># 504 70: SUB LIMINAL PERCEPTION

'I># 55990:

VISUAL STIMULATION

I figure

I

4

psychological abstracts 127

SUBJECT 1 (BH)

___

1

PR08L£M

,_ ____

2

3

SOLVING

SUBJECT 4 (BP)

--------

VERBAL

I

5

RECOGNITION

---

1

4

7

2

ATTENTION A

PROBLEM

SHADOWING

SOLVING

3

5

4

8

PERCEPTUAL PERCEPTUAL LEARNING DEVELOPMENT

SUBJECT 2 (CH)

7 RECOGNITION LEARNING

SUBJECT 5 (LP)

, ....

N (X)

1

PROBL£11

2

&OLVING

e

3

4

7

5

)

1

:z

..

3

8

RECOGNITION OF

INCIDENTAL

PROBLEM

ATTENTIONAL

INDIVIDUAL DIFFERENCES

VERBAL STIMULI

LEARNING

SOLVING

LEARNING

IN RECOGNITION

Win (.4, .2) 12 CODES OR

SUBJECT 3 (PG)

VISUAL

DESCRIPTORS PROBLEM --�

SOLVING

LEARNING

WORD RECOGNITION

RECOGNITION

VISUAL DEVELOPM£NTAL

2

1

I

3

PAOBUM

VISUAL

SOLVING

RECOGNITION

7

5 LEARNING

1

2

(PROBLEM SOLVING)

3

I (WORD)

4

7

5 (NIL)

N

• S

FIGURE 5

-------------------